[libside.git] / src / side.c

// SPDX-License-Identifier: MIT
/*
 * Copyright 2022 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 */

#include <side/trace.h>
#include <string.h>
#include <assert.h>

#include "rcu.h"
#include "list.h"
#include "rculist.h"

/* Top 8 bits reserved for kernel tracer use. */
#if SIDE_BITS_PER_LONG == 64
# define SIDE_EVENT_ENABLED_KERNEL_MASK			0xFF00000000000000ULL
# define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK 	0x8000000000000000ULL

/* Allow 2^56 tracer references on an event. */
# define SIDE_EVENT_ENABLED_USER_MASK			0x00FFFFFFFFFFFFFFULL
#else
# define SIDE_EVENT_ENABLED_KERNEL_MASK			0xFF000000UL
# define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK	0x80000000UL

/* Allow 2^24 tracer references on an event. */
# define SIDE_EVENT_ENABLED_USER_MASK			0x00FFFFFFUL
#endif

struct side_events_register_handle {
	struct side_list_node node;
	struct side_event_description **events;
	uint32_t nr_events;
};

struct side_tracer_handle {
	struct side_list_node node;
	void (*cb)(enum side_tracer_notification notif,
		struct side_event_description **events, uint32_t nr_events, void *priv);
	void *priv;
};

struct side_statedump_request_handle {
	struct side_list_node node;	/* RCU list. */
	void (*cb)(void);
};

struct side_callback {
	union {
		void (*call)(const struct side_event_description *desc,
			const struct side_arg_vec *side_arg_vec,
			void *priv);
		void (*call_variadic)(const struct side_event_description *desc,
			const struct side_arg_vec *side_arg_vec,
			const struct side_arg_dynamic_struct *var_struct,
			void *priv);
	} u;
	void *priv;
	void *key;
};

static struct side_rcu_gp_state event_rcu_gp, statedump_rcu_gp;

/*
 * Lazy initialization for early use within library constructors.
 */
static bool initialized;
/*
 * Do not register/unregister any more events after destructor.
 */
static bool finalized;

/*
 * Recursive mutex to allow tracer callbacks to use the side API.
 */
static pthread_mutex_t side_event_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
static pthread_mutex_t side_statedump_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;

static DEFINE_SIDE_LIST_HEAD(side_events_list);
static DEFINE_SIDE_LIST_HEAD(side_tracer_list);
static DEFINE_SIDE_LIST_HEAD(side_statedump_list);

/*
 * Callback filter key for state dump.
 */
static __thread void *filter_key;

/*
 * The empty callback has a NULL function callback pointer, which stops
 * iteration on the array of callbacks immediately.
 */
const char side_empty_callback[sizeof(struct side_callback)];

static
void _side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec, void *key)
{
	struct side_rcu_read_state rcu_read_state;
	const struct side_event_state_0 *es0;
	const struct side_callback *side_cb;
	uintptr_t enabled;

	if (side_unlikely(finalized))
		return;
	if (side_unlikely(!initialized))
		side_init();
	if (side_unlikely(event_state->version != 0))
		abort();
	es0 = side_container_of(event_state, const struct side_event_state_0, parent);
	assert(!(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC));
	enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
	if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) {
		// TODO: call kernel write.
	}
	side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
	for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call != NULL; side_cb++) {
		/* A NULL key is always a match. */
		if (key && side_cb->key && side_cb->key != key)
			continue;
		side_cb->u.call(es0->desc, side_arg_vec, side_cb->priv);
	}
	side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
}

void side_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec)
{
	_side_call(event_state, side_arg_vec, NULL);
}

void side_statedump_call(const struct side_event_state *event_state, const struct side_arg_vec *side_arg_vec)
{
	_side_call(event_state, side_arg_vec, filter_key);
}

static
void _side_call_variadic(const struct side_event_state *event_state,
	const struct side_arg_vec *side_arg_vec,
	const struct side_arg_dynamic_struct *var_struct,
	void *key)
{
	struct side_rcu_read_state rcu_read_state;
	const struct side_event_state_0 *es0;
	const struct side_callback *side_cb;
	uintptr_t enabled;

	if (side_unlikely(finalized))
		return;
	if (side_unlikely(!initialized))
		side_init();
	if (side_unlikely(event_state->version != 0))
		abort();
	es0 = side_container_of(event_state, const struct side_event_state_0, parent);
	assert(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC);
	enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
	if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) {
		// TODO: call kernel write.
	}
	side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
	for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call_variadic != NULL; side_cb++) {
		/* A NULL key is always a match. */
		if (key && side_cb->key && side_cb->key != key)
			continue;
		side_cb->u.call_variadic(es0->desc, side_arg_vec, var_struct, side_cb->priv);
	}
	side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
}

void side_call_variadic(const struct side_event_state *event_state,
	const struct side_arg_vec *side_arg_vec,
	const struct side_arg_dynamic_struct *var_struct)
{
	_side_call_variadic(event_state, side_arg_vec, var_struct, NULL);
}

void side_statedump_call_variadic(const struct side_event_state *event_state,
	const struct side_arg_vec *side_arg_vec,
	const struct side_arg_dynamic_struct *var_struct)
{
	_side_call_variadic(event_state, side_arg_vec, var_struct, filter_key);
}

static
const struct side_callback *side_tracer_callback_lookup(
		const struct side_event_description *desc,
		void *call, void *priv, void *key)
{
	struct side_event_state *event_state = side_ptr_get(desc->state);
	const struct side_event_state_0 *es0;
	const struct side_callback *cb;

	if (side_unlikely(event_state->version != 0))
		abort();
	es0 = side_container_of(event_state, const struct side_event_state_0, parent);
	for (cb = es0->callbacks; cb->u.call != NULL; cb++) {
		if ((void *) cb->u.call == call && cb->priv == priv && cb->key == key)
			return cb;
	}
	return NULL;
}

static
int _side_tracer_callback_register(struct side_event_description *desc,
		void *call, void *priv, void *key)
{
	struct side_event_state *event_state;
	struct side_callback *old_cb, *new_cb;
	struct side_event_state_0 *es0;
	int ret = SIDE_ERROR_OK;
	uint32_t old_nr_cb;

	if (!call)
		return SIDE_ERROR_INVAL;
	if (finalized)
		return SIDE_ERROR_EXITING;
	if (!initialized)
		side_init();
	pthread_mutex_lock(&side_event_lock);
	event_state = side_ptr_get(desc->state);
	if (side_unlikely(event_state->version != 0))
		abort();
	es0 = side_container_of(event_state, struct side_event_state_0, parent);
	old_nr_cb = es0->nr_callbacks;
	if (old_nr_cb == UINT32_MAX) {
		ret = SIDE_ERROR_INVAL;
		goto unlock;
	}
	/* Reject duplicate (call, priv) tuples. */
	if (side_tracer_callback_lookup(desc, call, priv, key)) {
		ret = SIDE_ERROR_EXIST;
		goto unlock;
	}
	old_cb = (struct side_callback *) es0->callbacks;
	/* old_nr_cb + 1 (new cb) + 1 (NULL) */
	new_cb = (struct side_callback *) calloc(old_nr_cb + 2, sizeof(struct side_callback));
	if (!new_cb) {
		ret = SIDE_ERROR_NOMEM;
		goto unlock;
	}
	memcpy(new_cb, old_cb, old_nr_cb);
	if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
		new_cb[old_nr_cb].u.call_variadic =
			(side_tracer_callback_variadic_func) call;
	else
		new_cb[old_nr_cb].u.call =
			(side_tracer_callback_func) call;
	new_cb[old_nr_cb].priv = priv;
	new_cb[old_nr_cb].key = key;
	/* High order bits are already zeroed. */
	side_rcu_assign_pointer(es0->callbacks, new_cb);
	side_rcu_wait_grace_period(&event_rcu_gp);
	if (old_nr_cb)
		free(old_cb);
	es0->nr_callbacks++;
	/* Increment concurrently with kernel setting the top bits. */
	if (!old_nr_cb)
		(void) __atomic_add_fetch(&es0->enabled, 1, __ATOMIC_RELAXED);
unlock:
	pthread_mutex_unlock(&side_event_lock);
	return ret;
}

int side_tracer_callback_register(struct side_event_description *desc,
		side_tracer_callback_func call,
		void *priv, void *key)
{
	if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
		return SIDE_ERROR_INVAL;
	return _side_tracer_callback_register(desc, (void *) call, priv, key);
}

int side_tracer_callback_variadic_register(struct side_event_description *desc,
		side_tracer_callback_variadic_func call_variadic,
		void *priv, void *key)
{
	if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
		return SIDE_ERROR_INVAL;
	return _side_tracer_callback_register(desc, (void *) call_variadic, priv, key);
}

static int _side_tracer_callback_unregister(struct side_event_description *desc,
		void *call, void *priv, void *key)
{
	struct side_event_state *event_state;
	struct side_callback *old_cb, *new_cb;
	const struct side_callback *cb_pos;
	struct side_event_state_0 *es0;
	uint32_t pos_idx;
	int ret = SIDE_ERROR_OK;
	uint32_t old_nr_cb;

	if (!call)
		return SIDE_ERROR_INVAL;
	if (finalized)
		return SIDE_ERROR_EXITING;
	if (!initialized)
		side_init();
	pthread_mutex_lock(&side_event_lock);
	event_state = side_ptr_get(desc->state);
	if (side_unlikely(event_state->version != 0))
		abort();
	es0 = side_container_of(event_state, struct side_event_state_0, parent);
	cb_pos = side_tracer_callback_lookup(desc, call, priv, key);
	if (!cb_pos) {
		ret = SIDE_ERROR_NOENT;
		goto unlock;
	}
	old_nr_cb = es0->nr_callbacks;
	old_cb = (struct side_callback *) es0->callbacks;
	if (old_nr_cb == 1) {
		new_cb = (struct side_callback *) &side_empty_callback;
	} else {
		pos_idx = cb_pos - es0->callbacks;
		/* Remove entry at pos_idx. */
		/* old_nr_cb - 1 (removed cb) + 1 (NULL) */
		new_cb = (struct side_callback *) calloc(old_nr_cb, sizeof(struct side_callback));
		if (!new_cb) {
			ret = SIDE_ERROR_NOMEM;
			goto unlock;
		}
		memcpy(new_cb, old_cb, pos_idx);
		memcpy(&new_cb[pos_idx], &old_cb[pos_idx + 1], old_nr_cb - pos_idx - 1);
	}
	/* High order bits are already zeroed. */
	side_rcu_assign_pointer(es0->callbacks, new_cb);
	side_rcu_wait_grace_period(&event_rcu_gp);
	free(old_cb);
	es0->nr_callbacks--;
	/* Decrement concurrently with kernel setting the top bits. */
	if (old_nr_cb == 1)
		(void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
unlock:
	pthread_mutex_unlock(&side_event_lock);
	return ret;
}

int side_tracer_callback_unregister(struct side_event_description *desc,
		side_tracer_callback_func call,
		void *priv, void *key)
{
	if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
		return SIDE_ERROR_INVAL;
	return _side_tracer_callback_unregister(desc, (void *) call, priv, key);
}

int side_tracer_callback_variadic_unregister(struct side_event_description *desc,
		side_tracer_callback_variadic_func call_variadic,
		void *priv, void *key)
{
	if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
		return SIDE_ERROR_INVAL;
	return _side_tracer_callback_unregister(desc, (void *) call_variadic, priv, key);
}

struct side_events_register_handle *side_events_register(struct side_event_description **events, uint32_t nr_events)
{
	struct side_events_register_handle *events_handle = NULL;
	struct side_tracer_handle *tracer_handle;

	if (finalized)
		return NULL;
	if (!initialized)
		side_init();
	events_handle = (struct side_events_register_handle *)
			calloc(1, sizeof(struct side_events_register_handle));
	if (!events_handle)
		return NULL;
	events_handle->events = events;
	events_handle->nr_events = nr_events;

	pthread_mutex_lock(&side_event_lock);
	side_list_insert_node_tail(&side_events_list, &events_handle->node);
	side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
		tracer_handle->cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
			events, nr_events, tracer_handle->priv);
	}
	pthread_mutex_unlock(&side_event_lock);
	//TODO: call event batch register ioctl
	return events_handle;
}

static
void side_event_remove_callbacks(struct side_event_description *desc)
{
	struct side_event_state *event_state = side_ptr_get(desc->state);
	struct side_event_state_0 *es0;
	struct side_callback *old_cb;
	uint32_t nr_cb;

	if (side_unlikely(event_state->version != 0))
		abort();
	es0 = side_container_of(event_state, struct side_event_state_0, parent);
	nr_cb = es0->nr_callbacks;
	if (!nr_cb)
		return;
	old_cb = (struct side_callback *) es0->callbacks;
	(void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
	/*
	 * Setting the state back to 0 cb and empty callbacks out of
	 * caution. This should not matter because instrumentation is
	 * unreachable.
	 */
	es0->nr_callbacks = 0;
	side_rcu_assign_pointer(es0->callbacks, &side_empty_callback);
	/*
	 * No need to wait for grace period because instrumentation is
	 * unreachable.
	 */
	free(old_cb);
}

/*
 * Unregister event handle. At this point, all side events in that
 * handle should be unreachable.
 */
void side_events_unregister(struct side_events_register_handle *events_handle)
{
	struct side_tracer_handle *tracer_handle;
	uint32_t i;

	if (!events_handle)
		return;
	if (finalized)
		return;
	if (!initialized)
		side_init();
	pthread_mutex_lock(&side_event_lock);
	side_list_remove_node(&events_handle->node);
	side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
		tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
			events_handle->events, events_handle->nr_events,
			tracer_handle->priv);
	}
	for (i = 0; i < events_handle->nr_events; i++) {
		struct side_event_description *event = events_handle->events[i];

		/* Skip NULL pointers */
		if (!event)
			continue;
		side_event_remove_callbacks(event);
	}
	pthread_mutex_unlock(&side_event_lock);
	//TODO: call event batch unregister ioctl
	free(events_handle);
}

struct side_tracer_handle *side_tracer_event_notification_register(
		void (*cb)(enum side_tracer_notification notif,
			struct side_event_description **events, uint32_t nr_events, void *priv),
		void *priv)
{
	struct side_tracer_handle *tracer_handle;
	struct side_events_register_handle *events_handle;

	if (finalized)
		return NULL;
	if (!initialized)
		side_init();
	tracer_handle = (struct side_tracer_handle *)
				calloc(1, sizeof(struct side_tracer_handle));
	if (!tracer_handle)
		return NULL;
	pthread_mutex_lock(&side_event_lock);
	tracer_handle->cb = cb;
	tracer_handle->priv = priv;
	side_list_insert_node_tail(&side_tracer_list, &tracer_handle->node);
	side_list_for_each_entry(events_handle, &side_events_list, node) {
		cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
			events_handle->events, events_handle->nr_events, priv);
	}
	pthread_mutex_unlock(&side_event_lock);
	return tracer_handle;
}

void side_tracer_event_notification_unregister(struct side_tracer_handle *tracer_handle)
{
	struct side_events_register_handle *events_handle;

	if (finalized)
		return;
	if (!initialized)
		side_init();
	pthread_mutex_lock(&side_event_lock);
	side_list_for_each_entry(events_handle, &side_events_list, node) {
		tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
			events_handle->events, events_handle->nr_events,
			tracer_handle->priv);
	}
	side_list_remove_node(&tracer_handle->node);
	pthread_mutex_unlock(&side_event_lock);
	free(tracer_handle);
}

struct side_statedump_request_handle *side_statedump_request_notification_register(void (*statedump_cb)(void))
{
	struct side_statedump_request_handle *handle;

	if (finalized)
		return NULL;
	if (!initialized)
		side_init();
	/*
	 * The statedump request notification should not be registered
	 * from a notification callback.
	 */
	assert(filter_key == NULL);
	handle = (struct side_statedump_request_handle *)
				calloc(1, sizeof(struct side_statedump_request_handle));
	if (!handle)
		return NULL;
	handle->cb = statedump_cb;

	pthread_mutex_lock(&side_statedump_lock);
	side_list_insert_node_tail_rcu(&side_statedump_list, &handle->node);
	pthread_mutex_unlock(&side_statedump_lock);

	/* Invoke callback for all tracers. */
	statedump_cb();

	return handle;
}

void side_statedump_request_notification_unregister(struct side_statedump_request_handle *handle)
{
	if (finalized)
		return;
	if (!initialized)
		side_init();
	assert(filter_key == NULL);

	pthread_mutex_lock(&side_statedump_lock);
	side_list_remove_node_rcu(&handle->node);
	pthread_mutex_unlock(&side_statedump_lock);

	side_rcu_wait_grace_period(&statedump_rcu_gp);
	free(handle);
}

void side_tracer_statedump_request(void *key)
{
	struct side_statedump_request_handle *handle;
	struct side_rcu_read_state rcu_read_state;

	/* Invoke the state dump callback specifically for the tracer key. */
	filter_key = key;
	side_rcu_read_begin(&statedump_rcu_gp, &rcu_read_state);
	side_list_for_each_entry_rcu(handle, &side_statedump_list, node)
		handle->cb();
	side_rcu_read_end(&statedump_rcu_gp, &rcu_read_state);
	filter_key = NULL;
}

void side_init(void)
{
	if (initialized)
		return;
	side_rcu_gp_init(&event_rcu_gp);
	side_rcu_gp_init(&statedump_rcu_gp);
	initialized = true;
}

/*
 * side_exit() is executed from a library destructor. It can be called
 * explicitly at application exit as well. Concurrent side API use is
 * not expected at that point.
 */
void side_exit(void)
{
	struct side_events_register_handle *handle, *tmp;

	if (finalized)
		return;
	side_list_for_each_entry_safe(handle, tmp, &side_events_list, node)
		side_events_unregister(handle);
	side_rcu_gp_exit(&event_rcu_gp);
	side_rcu_gp_exit(&statedump_rcu_gp);
	finalized = true;
}
Commit	Line	Data
67337c4a MD	1	// SPDX-License-Identifier: MIT
	2	/*
	3	* Copyright 2022 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	4	*/
	5
	6	#include <side/trace.h>
	7	#include <string.h>
b1bf768c	8	#include <assert.h>
67337c4a MD	9
	10	#include "rcu.h"
	11	#include "list.h"
873bbf16	12	#include "rculist.h"
67337c4a MD	13
	14	/* Top 8 bits reserved for kernel tracer use. */
	15	#if SIDE_BITS_PER_LONG == 64
	16	# define SIDE_EVENT_ENABLED_KERNEL_MASK 0xFF00000000000000ULL
	17	# define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK 0x8000000000000000ULL
	18
	19	/* Allow 2^56 tracer references on an event. */
	20	# define SIDE_EVENT_ENABLED_USER_MASK 0x00FFFFFFFFFFFFFFULL
	21	#else
	22	# define SIDE_EVENT_ENABLED_KERNEL_MASK 0xFF000000UL
	23	# define SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK 0x80000000UL
	24
	25	/* Allow 2^24 tracer references on an event. */
	26	# define SIDE_EVENT_ENABLED_USER_MASK 0x00FFFFFFUL
	27	#endif
	28
	29	struct side_events_register_handle {
	30	struct side_list_node node;
	31	struct side_event_description **events;
	32	uint32_t nr_events;
	33	};
	34
	35	struct side_tracer_handle {
	36	struct side_list_node node;
	37	void (*cb)(enum side_tracer_notification notif,
	38	struct side_event_description *events, uint32_t nr_events, void priv);
	39	void *priv;
	40	};
	41
f0b01832	42	struct side_statedump_request_handle {
873bbf16	43	struct side_list_node node; /* RCU list. */
f0b01832 MD	44	void (*cb)(void);
	45	};
	46
867b4725 MD	47	struct side_callback {
	48	union {
	49	void (call)(const struct side_event_description desc,
	50	const struct side_arg_vec *side_arg_vec,
	51	void *priv);
	52	void (call_variadic)(const struct side_event_description desc,
	53	const struct side_arg_vec *side_arg_vec,
	54	const struct side_arg_dynamic_struct *var_struct,
	55	void *priv);
	56	} u;
	57	void *priv;
92c377f9	58	void *key;
867b4725 MD	59	};
867b4725 MD	60
873bbf16	61	static struct side_rcu_gp_state event_rcu_gp, statedump_rcu_gp;
67337c4a MD	62
	63	/*
	64	* Lazy initialization for early use within library constructors.
	65	*/
	66	static bool initialized;
	67	/*
	68	* Do not register/unregister any more events after destructor.
	69	*/
	70	static bool finalized;
	71
	72	/*
	73	* Recursive mutex to allow tracer callbacks to use the side API.
	74	*/
873bbf16 MD	75	static pthread_mutex_t side_event_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
873bbf16 MD	76	static pthread_mutex_t side_statedump_lock = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
67337c4a MD	77
	78	static DEFINE_SIDE_LIST_HEAD(side_events_list);
	79	static DEFINE_SIDE_LIST_HEAD(side_tracer_list);
f0b01832	80	static DEFINE_SIDE_LIST_HEAD(side_statedump_list);
67337c4a	81
74be90b7 MD	82	/*
	83	* Callback filter key for state dump.
	84	*/
	85	static __thread void *filter_key;
	86
67337c4a MD	87	/*
	88	* The empty callback has a NULL function callback pointer, which stops
	89	* iteration on the array of callbacks immediately.
	90	*/
867b4725	91	const char side_empty_callback[sizeof(struct side_callback)];
67337c4a	92
74be90b7 MD	93	static
74be90b7 MD	94	void _side_call(const struct side_event_state event_state, const struct side_arg_vec side_arg_vec, void *key)
67337c4a MD	95	{
67337c4a MD	96	struct side_rcu_read_state rcu_read_state;
b2a84b9f	97	const struct side_event_state_0 *es0;
67337c4a MD	98	const struct side_callback *side_cb;
	99	uintptr_t enabled;
	100
	101	if (side_unlikely(finalized))
	102	return;
	103	if (side_unlikely(!initialized))
	104	side_init();
b2a84b9f MD	105	if (side_unlikely(event_state->version != 0))
b2a84b9f MD	106	abort();
49aea3ef	107	es0 = side_container_of(event_state, const struct side_event_state_0, parent);
7269a8a3	108	assert(!(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC));
b2a84b9f	109	enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
67337c4a MD	110	if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) {
	111	// TODO: call kernel write.
	112	}
873bbf16	113	side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
92c377f9 MD	114	for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call != NULL; side_cb++) {
	115	/* A NULL key is always a match. */
	116	if (key && side_cb->key && side_cb->key != key)
	117	continue;
7269a8a3	118	side_cb->u.call(es0->desc, side_arg_vec, side_cb->priv);
92c377f9	119	}
873bbf16	120	side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
67337c4a MD	121	}
67337c4a MD	122
92c377f9 MD	123	void side_call(const struct side_event_state event_state, const struct side_arg_vec side_arg_vec)
92c377f9 MD	124	{
74be90b7	125	_side_call(event_state, side_arg_vec, NULL);
92c377f9 MD	126	}
92c377f9 MD	127
f0b01832	128	void side_statedump_call(const struct side_event_state event_state, const struct side_arg_vec side_arg_vec)
74be90b7 MD	129	{
	130	_side_call(event_state, side_arg_vec, filter_key);
	131	}
	132
	133	static
	134	void _side_call_variadic(const struct side_event_state *event_state,
67337c4a	135	const struct side_arg_vec *side_arg_vec,
92c377f9 MD	136	const struct side_arg_dynamic_struct *var_struct,
92c377f9 MD	137	void *key)
67337c4a MD	138	{
67337c4a MD	139	struct side_rcu_read_state rcu_read_state;
b2a84b9f	140	const struct side_event_state_0 *es0;
67337c4a MD	141	const struct side_callback *side_cb;
	142	uintptr_t enabled;
	143
	144	if (side_unlikely(finalized))
	145	return;
	146	if (side_unlikely(!initialized))
	147	side_init();
b2a84b9f MD	148	if (side_unlikely(event_state->version != 0))
b2a84b9f MD	149	abort();
49aea3ef	150	es0 = side_container_of(event_state, const struct side_event_state_0, parent);
7269a8a3	151	assert(es0->desc->flags & SIDE_EVENT_FLAG_VARIADIC);
b2a84b9f	152	enabled = __atomic_load_n(&es0->enabled, __ATOMIC_RELAXED);
67337c4a MD	153	if (side_unlikely(enabled & SIDE_EVENT_ENABLED_KERNEL_USER_EVENT_MASK)) {
	154	// TODO: call kernel write.
	155	}
873bbf16	156	side_rcu_read_begin(&event_rcu_gp, &rcu_read_state);
92c377f9 MD	157	for (side_cb = side_rcu_dereference(es0->callbacks); side_cb->u.call_variadic != NULL; side_cb++) {
	158	/* A NULL key is always a match. */
	159	if (key && side_cb->key && side_cb->key != key)
	160	continue;
7269a8a3	161	side_cb->u.call_variadic(es0->desc, side_arg_vec, var_struct, side_cb->priv);
92c377f9	162	}
873bbf16	163	side_rcu_read_end(&event_rcu_gp, &rcu_read_state);
67337c4a MD	164	}
67337c4a MD	165
92c377f9 MD	166	void side_call_variadic(const struct side_event_state *event_state,
	167	const struct side_arg_vec *side_arg_vec,
	168	const struct side_arg_dynamic_struct *var_struct)
	169	{
74be90b7 MD	170	_side_call_variadic(event_state, side_arg_vec, var_struct, NULL);
	171	}
	172
f0b01832	173	void side_statedump_call_variadic(const struct side_event_state *event_state,
74be90b7 MD	174	const struct side_arg_vec *side_arg_vec,
	175	const struct side_arg_dynamic_struct *var_struct)
	176	{
	177	_side_call_variadic(event_state, side_arg_vec, var_struct, filter_key);
92c377f9 MD	178	}
92c377f9 MD	179
67337c4a MD	180	static
	181	const struct side_callback *side_tracer_callback_lookup(
	182	const struct side_event_description *desc,
92c377f9	183	void call, void priv, void *key)
67337c4a	184	{
0b9e59d6	185	struct side_event_state *event_state = side_ptr_get(desc->state);
b2a84b9f	186	const struct side_event_state_0 *es0;
67337c4a MD	187	const struct side_callback *cb;
67337c4a MD	188
b2a84b9f MD	189	if (side_unlikely(event_state->version != 0))
b2a84b9f MD	190	abort();
49aea3ef	191	es0 = side_container_of(event_state, const struct side_event_state_0, parent);
7269a8a3	192	for (cb = es0->callbacks; cb->u.call != NULL; cb++) {
92c377f9	193	if ((void *) cb->u.call == call && cb->priv == priv && cb->key == key)
67337c4a MD	194	return cb;
	195	}
	196	return NULL;
	197	}
	198
	199	static
	200	int _side_tracer_callback_register(struct side_event_description *desc,
92c377f9	201	void call, void priv, void *key)
67337c4a	202	{
0b9e59d6	203	struct side_event_state *event_state;
67337c4a	204	struct side_callback old_cb, new_cb;
b2a84b9f	205	struct side_event_state_0 *es0;
67337c4a MD	206	int ret = SIDE_ERROR_OK;
	207	uint32_t old_nr_cb;
	208
	209	if (!call)
	210	return SIDE_ERROR_INVAL;
	211	if (finalized)
	212	return SIDE_ERROR_EXITING;
	213	if (!initialized)
	214	side_init();
873bbf16	215	pthread_mutex_lock(&side_event_lock);
0b9e59d6	216	event_state = side_ptr_get(desc->state);
b2a84b9f MD	217	if (side_unlikely(event_state->version != 0))
b2a84b9f MD	218	abort();
49aea3ef	219	es0 = side_container_of(event_state, struct side_event_state_0, parent);
3cac1780	220	old_nr_cb = es0->nr_callbacks;
67337c4a MD	221	if (old_nr_cb == UINT32_MAX) {
	222	ret = SIDE_ERROR_INVAL;
	223	goto unlock;
	224	}
	225	/* Reject duplicate (call, priv) tuples. */
92c377f9	226	if (side_tracer_callback_lookup(desc, call, priv, key)) {
67337c4a MD	227	ret = SIDE_ERROR_EXIST;
	228	goto unlock;
	229	}
7269a8a3	230	old_cb = (struct side_callback *) es0->callbacks;
67337c4a MD	231	/* old_nr_cb + 1 (new cb) + 1 (NULL) */
	232	new_cb = (struct side_callback *) calloc(old_nr_cb + 2, sizeof(struct side_callback));
	233	if (!new_cb) {
	234	ret = SIDE_ERROR_NOMEM;
	235	goto unlock;
	236	}
	237	memcpy(new_cb, old_cb, old_nr_cb);
	238	if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
	239	new_cb[old_nr_cb].u.call_variadic =
	240	(side_tracer_callback_variadic_func) call;
	241	else
	242	new_cb[old_nr_cb].u.call =
	243	(side_tracer_callback_func) call;
	244	new_cb[old_nr_cb].priv = priv;
92c377f9	245	new_cb[old_nr_cb].key = key;
f60d8121	246	/* High order bits are already zeroed. */
7269a8a3	247	side_rcu_assign_pointer(es0->callbacks, new_cb);
873bbf16	248	side_rcu_wait_grace_period(&event_rcu_gp);
67337c4a MD	249	if (old_nr_cb)
67337c4a MD	250	free(old_cb);
3cac1780	251	es0->nr_callbacks++;
67337c4a MD	252	/* Increment concurrently with kernel setting the top bits. */
67337c4a MD	253	if (!old_nr_cb)
b2a84b9f	254	(void) __atomic_add_fetch(&es0->enabled, 1, __ATOMIC_RELAXED);
67337c4a	255	unlock:
873bbf16	256	pthread_mutex_unlock(&side_event_lock);
67337c4a MD	257	return ret;
	258	}
	259
	260	int side_tracer_callback_register(struct side_event_description *desc,
	261	side_tracer_callback_func call,
92c377f9	262	void priv, void key)
67337c4a MD	263	{
	264	if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
	265	return SIDE_ERROR_INVAL;
92c377f9	266	return _side_tracer_callback_register(desc, (void *) call, priv, key);
67337c4a MD	267	}
	268
	269	int side_tracer_callback_variadic_register(struct side_event_description *desc,
	270	side_tracer_callback_variadic_func call_variadic,
92c377f9	271	void priv, void key)
67337c4a MD	272	{
	273	if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
	274	return SIDE_ERROR_INVAL;
92c377f9	275	return _side_tracer_callback_register(desc, (void *) call_variadic, priv, key);
67337c4a MD	276	}
	277
	278	static int _side_tracer_callback_unregister(struct side_event_description *desc,
92c377f9	279	void call, void priv, void *key)
67337c4a	280	{
0b9e59d6	281	struct side_event_state *event_state;
67337c4a MD	282	struct side_callback old_cb, new_cb;
67337c4a MD	283	const struct side_callback *cb_pos;
b2a84b9f	284	struct side_event_state_0 *es0;
67337c4a MD	285	uint32_t pos_idx;
	286	int ret = SIDE_ERROR_OK;
	287	uint32_t old_nr_cb;
	288
	289	if (!call)
	290	return SIDE_ERROR_INVAL;
	291	if (finalized)
	292	return SIDE_ERROR_EXITING;
	293	if (!initialized)
	294	side_init();
873bbf16	295	pthread_mutex_lock(&side_event_lock);
0b9e59d6	296	event_state = side_ptr_get(desc->state);
b2a84b9f MD	297	if (side_unlikely(event_state->version != 0))
b2a84b9f MD	298	abort();
49aea3ef	299	es0 = side_container_of(event_state, struct side_event_state_0, parent);
92c377f9	300	cb_pos = side_tracer_callback_lookup(desc, call, priv, key);
67337c4a MD	301	if (!cb_pos) {
	302	ret = SIDE_ERROR_NOENT;
	303	goto unlock;
	304	}
3cac1780	305	old_nr_cb = es0->nr_callbacks;
7269a8a3	306	old_cb = (struct side_callback *) es0->callbacks;
67337c4a MD	307	if (old_nr_cb == 1) {
	308	new_cb = (struct side_callback *) &side_empty_callback;
	309	} else {
7269a8a3	310	pos_idx = cb_pos - es0->callbacks;
67337c4a MD	311	/* Remove entry at pos_idx. */
	312	/* old_nr_cb - 1 (removed cb) + 1 (NULL) */
	313	new_cb = (struct side_callback *) calloc(old_nr_cb, sizeof(struct side_callback));
	314	if (!new_cb) {
	315	ret = SIDE_ERROR_NOMEM;
	316	goto unlock;
	317	}
	318	memcpy(new_cb, old_cb, pos_idx);
	319	memcpy(&new_cb[pos_idx], &old_cb[pos_idx + 1], old_nr_cb - pos_idx - 1);
	320	}
f60d8121	321	/* High order bits are already zeroed. */
7269a8a3	322	side_rcu_assign_pointer(es0->callbacks, new_cb);
873bbf16	323	side_rcu_wait_grace_period(&event_rcu_gp);
67337c4a	324	free(old_cb);
3cac1780	325	es0->nr_callbacks--;
67337c4a MD	326	/* Decrement concurrently with kernel setting the top bits. */
67337c4a MD	327	if (old_nr_cb == 1)
b2a84b9f	328	(void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
67337c4a	329	unlock:
873bbf16	330	pthread_mutex_unlock(&side_event_lock);
67337c4a MD	331	return ret;
	332	}
	333
	334	int side_tracer_callback_unregister(struct side_event_description *desc,
	335	side_tracer_callback_func call,
92c377f9	336	void priv, void key)
67337c4a MD	337	{
	338	if (desc->flags & SIDE_EVENT_FLAG_VARIADIC)
	339	return SIDE_ERROR_INVAL;
92c377f9	340	return _side_tracer_callback_unregister(desc, (void *) call, priv, key);
67337c4a MD	341	}
	342
	343	int side_tracer_callback_variadic_unregister(struct side_event_description *desc,
	344	side_tracer_callback_variadic_func call_variadic,
92c377f9	345	void priv, void key)
67337c4a MD	346	{
	347	if (!(desc->flags & SIDE_EVENT_FLAG_VARIADIC))
	348	return SIDE_ERROR_INVAL;
92c377f9	349	return _side_tracer_callback_unregister(desc, (void *) call_variadic, priv, key);
67337c4a MD	350	}
	351
	352	struct side_events_register_handle side_events_register(struct side_event_description *events, uint32_t nr_events)
	353	{
	354	struct side_events_register_handle *events_handle = NULL;
	355	struct side_tracer_handle *tracer_handle;
	356
	357	if (finalized)
	358	return NULL;
	359	if (!initialized)
	360	side_init();
	361	events_handle = (struct side_events_register_handle *)
	362	calloc(1, sizeof(struct side_events_register_handle));
	363	if (!events_handle)
	364	return NULL;
	365	events_handle->events = events;
	366	events_handle->nr_events = nr_events;
	367
873bbf16	368	pthread_mutex_lock(&side_event_lock);
67337c4a MD	369	side_list_insert_node_tail(&side_events_list, &events_handle->node);
	370	side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
	371	tracer_handle->cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
	372	events, nr_events, tracer_handle->priv);
	373	}
873bbf16	374	pthread_mutex_unlock(&side_event_lock);
67337c4a MD	375	//TODO: call event batch register ioctl
	376	return events_handle;
	377	}
	378
	379	static
	380	void side_event_remove_callbacks(struct side_event_description *desc)
	381	{
0b9e59d6	382	struct side_event_state *event_state = side_ptr_get(desc->state);
b2a84b9f	383	struct side_event_state_0 *es0;
67337c4a	384	struct side_callback *old_cb;
3cac1780	385	uint32_t nr_cb;
67337c4a	386
b2a84b9f MD	387	if (side_unlikely(event_state->version != 0))
b2a84b9f MD	388	abort();
49aea3ef	389	es0 = side_container_of(event_state, struct side_event_state_0, parent);
3cac1780 MD	390	nr_cb = es0->nr_callbacks;
	391	if (!nr_cb)
	392	return;
7269a8a3	393	old_cb = (struct side_callback *) es0->callbacks;
b2a84b9f	394	(void) __atomic_add_fetch(&es0->enabled, -1, __ATOMIC_RELAXED);
67337c4a MD	395	/*
	396	* Setting the state back to 0 cb and empty callbacks out of
	397	* caution. This should not matter because instrumentation is
	398	* unreachable.
	399	*/
3cac1780	400	es0->nr_callbacks = 0;
7269a8a3	401	side_rcu_assign_pointer(es0->callbacks, &side_empty_callback);
67337c4a MD	402	/*
	403	* No need to wait for grace period because instrumentation is
	404	* unreachable.
	405	*/
	406	free(old_cb);
	407	}
	408
	409	/*
	410	* Unregister event handle. At this point, all side events in that
	411	* handle should be unreachable.
	412	*/
	413	void side_events_unregister(struct side_events_register_handle *events_handle)
	414	{
	415	struct side_tracer_handle *tracer_handle;
	416	uint32_t i;
	417
	418	if (!events_handle)
	419	return;
	420	if (finalized)
	421	return;
	422	if (!initialized)
	423	side_init();
873bbf16	424	pthread_mutex_lock(&side_event_lock);
67337c4a MD	425	side_list_remove_node(&events_handle->node);
	426	side_list_for_each_entry(tracer_handle, &side_tracer_list, node) {
	427	tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
	428	events_handle->events, events_handle->nr_events,
	429	tracer_handle->priv);
	430	}
	431	for (i = 0; i < events_handle->nr_events; i++) {
	432	struct side_event_description *event = events_handle->events[i];
	433
	434	/* Skip NULL pointers */
	435	if (!event)
	436	continue;
	437	side_event_remove_callbacks(event);
	438	}
873bbf16	439	pthread_mutex_unlock(&side_event_lock);
67337c4a MD	440	//TODO: call event batch unregister ioctl
	441	free(events_handle);
	442	}
	443
	444	struct side_tracer_handle *side_tracer_event_notification_register(
	445	void (*cb)(enum side_tracer_notification notif,
	446	struct side_event_description *events, uint32_t nr_events, void priv),
	447	void *priv)
	448	{
	449	struct side_tracer_handle *tracer_handle;
	450	struct side_events_register_handle *events_handle;
	451
	452	if (finalized)
	453	return NULL;
	454	if (!initialized)
	455	side_init();
	456	tracer_handle = (struct side_tracer_handle *)
	457	calloc(1, sizeof(struct side_tracer_handle));
	458	if (!tracer_handle)
	459	return NULL;
873bbf16	460	pthread_mutex_lock(&side_event_lock);
67337c4a MD	461	tracer_handle->cb = cb;
	462	tracer_handle->priv = priv;
	463	side_list_insert_node_tail(&side_tracer_list, &tracer_handle->node);
	464	side_list_for_each_entry(events_handle, &side_events_list, node) {
	465	cb(SIDE_TRACER_NOTIFICATION_INSERT_EVENTS,
	466	events_handle->events, events_handle->nr_events, priv);
	467	}
873bbf16	468	pthread_mutex_unlock(&side_event_lock);
67337c4a MD	469	return tracer_handle;
	470	}
	471
	472	void side_tracer_event_notification_unregister(struct side_tracer_handle *tracer_handle)
	473	{
	474	struct side_events_register_handle *events_handle;
	475
	476	if (finalized)
	477	return;
	478	if (!initialized)
	479	side_init();
873bbf16	480	pthread_mutex_lock(&side_event_lock);
67337c4a MD	481	side_list_for_each_entry(events_handle, &side_events_list, node) {
	482	tracer_handle->cb(SIDE_TRACER_NOTIFICATION_REMOVE_EVENTS,
	483	events_handle->events, events_handle->nr_events,
	484	tracer_handle->priv);
	485	}
	486	side_list_remove_node(&tracer_handle->node);
873bbf16	487	pthread_mutex_unlock(&side_event_lock);
be787080	488	free(tracer_handle);
67337c4a MD	489	}
67337c4a MD	490
f0b01832 MD	491	struct side_statedump_request_handle side_statedump_request_notification_register(void (statedump_cb)(void))
	492	{
	493	struct side_statedump_request_handle *handle;
	494
	495	if (finalized)
	496	return NULL;
	497	if (!initialized)
	498	side_init();
	499	/*
	500	* The statedump request notification should not be registered
	501	* from a notification callback.
	502	*/
	503	assert(filter_key == NULL);
	504	handle = (struct side_statedump_request_handle *)
	505	calloc(1, sizeof(struct side_statedump_request_handle));
	506	if (!handle)
	507	return NULL;
f0b01832	508	handle->cb = statedump_cb;
873bbf16 MD	509
	510	pthread_mutex_lock(&side_statedump_lock);
	511	side_list_insert_node_tail_rcu(&side_statedump_list, &handle->node);
	512	pthread_mutex_unlock(&side_statedump_lock);
	513
f0b01832 MD	514	/* Invoke callback for all tracers. */
f0b01832 MD	515	statedump_cb();
873bbf16	516
f0b01832 MD	517	return handle;
	518	}
	519
	520	void side_statedump_request_notification_unregister(struct side_statedump_request_handle *handle)
	521	{
	522	if (finalized)
	523	return;
	524	if (!initialized)
	525	side_init();
	526	assert(filter_key == NULL);
873bbf16 MD	527
	528	pthread_mutex_lock(&side_statedump_lock);
	529	side_list_remove_node_rcu(&handle->node);
	530	pthread_mutex_unlock(&side_statedump_lock);
	531
	532	side_rcu_wait_grace_period(&statedump_rcu_gp);
f0b01832 MD	533	free(handle);
	534	}
	535
	536	void side_tracer_statedump_request(void *key)
	537	{
	538	struct side_statedump_request_handle *handle;
873bbf16	539	struct side_rcu_read_state rcu_read_state;
f0b01832 MD	540
	541	/* Invoke the state dump callback specifically for the tracer key. */
	542	filter_key = key;
873bbf16 MD	543	side_rcu_read_begin(&statedump_rcu_gp, &rcu_read_state);
873bbf16 MD	544	side_list_for_each_entry_rcu(handle, &side_statedump_list, node)
f0b01832	545	handle->cb();
873bbf16	546	side_rcu_read_end(&statedump_rcu_gp, &rcu_read_state);
f0b01832 MD	547	filter_key = NULL;
	548	}
	549
67337c4a MD	550	void side_init(void)
	551	{
	552	if (initialized)
	553	return;
873bbf16 MD	554	side_rcu_gp_init(&event_rcu_gp);
873bbf16 MD	555	side_rcu_gp_init(&statedump_rcu_gp);
67337c4a MD	556	initialized = true;
	557	}
	558
	559	/*
	560	* side_exit() is executed from a library destructor. It can be called
	561	* explicitly at application exit as well. Concurrent side API use is
	562	* not expected at that point.
	563	*/
	564	void side_exit(void)
	565	{
	566	struct side_events_register_handle handle, tmp;
	567
	568	if (finalized)
	569	return;
	570	side_list_for_each_entry_safe(handle, tmp, &side_events_list, node)
	571	side_events_unregister(handle);
873bbf16 MD	572	side_rcu_gp_exit(&event_rcu_gp);
873bbf16 MD	573	side_rcu_gp_exit(&statedump_rcu_gp);
67337c4a MD	574	finalized = true;
67337c4a MD	575	}